SGMSE

SGMSE takes a different bet from the masking crowd: it treats enhancement as generation. A score-based diffusion model (an NCSN++ score network over the complex STFT) learns to reverse a stochastic process that starts at the noisy spectrogram and ends at clean speech, iteratively denoising over reverse-diffusion steps. The first version appeared at Interspeech 2022, followed by the SGMSE+ journal paper in TASLP 2023 extending it to joint enhancement and dereverberation, plus an ICASSP 2023 analysis comparing generative against discriminative approaches. The Hamburg group ships the official PyTorch implementation with pretrained checkpoints under MIT, and the same codebase underpins their EARS fullband benchmark dataset work. The cost of the generative approach is speed: iterative sampling is far slower than a single mask pass, so this is an offline or research model rather than a real-time one. Author-reported quality is competitive with the best discriminative models, and generative methods have a documented edge on badly degraded inputs where masking leaves artifacts. Independent replication is limited, which keeps the score honest.

• Papers (Interspeech 2022 / TASLP 2023, author-reported): competitive with state-of-the-art discriminative enhancers on enhancement and dereverberation; authors report generative models degrade more gracefully on severe corruption
• No independent benchmark replication found